|Número de publicación||US6370504 B1|
|Tipo de publicación||Concesión|
|Número de solicitud||US 09/083,361|
|Fecha de publicación||9 Abr 2002|
|Fecha de presentación||22 May 1998|
|Fecha de prioridad||29 May 1997|
|Número de publicación||083361, 09083361, US 6370504 B1, US 6370504B1, US-B1-6370504, US6370504 B1, US6370504B1|
|Inventores||Gregory L. Zick, Lawrence Yapp|
|Cesionario original||University Of Washington|
|Exportar cita||BiBTeX, EndNote, RefMan|
|Citas de patentes (3), Otras citas (1), Citada por (119), Clasificaciones (11), Eventos legales (9)|
|Enlaces externos: USPTO, Cesión de USPTO, Espacenet|
This application is based on prior copending provisional application Ser. No. 60/047,858, filed May 29, 1997, the benefit of the filing date of which is hereby claimed under 35 U.S.C. §119(e).
This invention generally relates to indexing files, and more specifically, pertains to the use of speech recognition of MPEG/Audio encoded files to permit indexing the files in regard to content.
The Motion Picture Experts Group (MPEG) set of standards have become an enabling technology for many of today's multimedia applications. MPEG encoded video facilitates the transmission of movies and other multimedia works over networks and the efficient storage of video data in a digital library. As video libraries for storing multimedia works become common and grow in size, the need to automatically and consistently index the content of videos becomes increasingly more important. A video that has been richly indexed can open a host of new applications. For example, knowing the location in a video where a favorite actor speaks or where a keyword is spoken can enable a variety of large scale implementations of hyperlinked video and facilitate new training and educational programs.
An audio track accompanying a video contains an enormous amount of information, some of which can be inferred by the human listener. For example, by listening to the audio portion independent of the video action, in addition to understanding what is being said, humans can infer who is doing the talking. Other types of information that can be inferred are gender of the speaker, language, emotion, and the style of the video material, e.g., dramatic or comedic. Therefore, a great opportunity lies in the automatic extraction of these audio elements so that they can be used as indices in framing queries and in effecting the retrieval of associated video segments.
There have been a great number of publications on work related to video indexing. The majority are focused on scene cuts and camera operations such as pan and zoom. Knowing only where a scene cut is or when a camera movement occurs in a video is not enough to develop a usable index of the video based on content. The practical application of indexing, particularly for non-professional applications, requires more content than has been provided by the prior art. Clearly, it is more likely that a home video user will employ an index to access a particular type of program in a video library rather than to locate a particular kind of camera shot.
Complete, general, and unrestricted automatic video indexing based on the analysis of the accompanying audio may not be possible for some time. Nevertheless, with restrictions, successful application of current speech recognition technology can be applied to automatic video indexing for domain specific applications. For example, in a sports video of a football game, a small set of keywords can be used and recorded onto the audio track to describe plays and formations such as “fake punt” and “I-32 wide right.” Being able to recognize these connected words in the accompanying audio can lead to a very useful index that allows a football coach to select desired plays and go directly to the associated video segment during a training session or analysis of the game recorded on the video.
From the foregoing discussion, it will be apparent that a technique capable of recognizing continuously uttered words in compressed MPEG/Audio for use as indices would be a valuable tool. Working in the compressed domain has a number of benefits. It saves computation time, since decompression is not necessary, and it saves disk space, since no intermediate file is created (e.g., as would be required if the audio track on the MPEG files had to be decompressed first).
There is not much published work on the direct analysis of the audio accompanying video material for speech recognition in video indexing. A system has been described to determine voiced, unvoiced, and silent classification directly on MPEG/Audio, and various time domain speech analysis techniques such as the short-time energy function for endpoint detection have been described in the prior art. These initial efforts contributed to simple video indexing based on audio, but the extracted information has been limited to the characteristics of the audio signal and not its content. However, just knowing where voicing and silence occurs is not enough. Additional information is needed in order to be useful for content-based video indices.
If the video contains a closed-captioned signal, an index can be readily made from the words extracted from the signal. A prior art system that makes use of this technique uses closed-captioned signal to generate pictorial transcripts of the video. Unfortunately, there is a problem with the limited amount of video material that comes with closed-captioning. In another prior art system, the Sphinx-II speech recognition system has been applied to a closed-captioned signal. However, when closed captioning was not available in this system, speech recognition was used exclusively. The reported word error rate ranged from ˜8-12% for speech benchmark evaluation to ˜85% for commercials.
Recently, an integrated approach of using image and speech analysis for video indexing has been described in the art. The video and audio are separated, with video encoded into AVI and audio into WAV files. A speech recognition system based on template matching was used in this system. However, using template matching for speech recognition is crude, as it has problems dealing with the variability that exists in speech.
Unlike the prior art described above, it would be preferable to derive speech features for recognition from the compressed MPEG/Audio domain, and to extract higher level indices that go beyond the characteristics of the audio signal by applying speech recognition. Accuracy of the process should be enhanced by using a robust technique and by restricting the scope for recognition to specific applications (i.e., not for general speech recognition).
It will be understood that MPEG-1 was used in developing the present invention, and as used throughout this description, the term “MPEG/Audio” will mean MPEG-1 and MPEG-2/Audio.) The underlying speech recognizer employed in a preferred form of the invention is based on the Hidden Markov model. Decompression is not required, as training and recognition is performed based on the extracted subbands of the MPEG/Audio files.
In accord with the present invention, a method is defined for recognizing speech in an MPEG file. The method includes the step of determining speech features of the MPEG file without first decompressing the MPEG file. Using the speech features, a speech recognition model is trained to recognize words in a different MPEG file. The speech recognition model is then employed to recognize words spoken in an audio portion of the different MPEG file, also without first decompressing that MPEG file.
The foregoing aspects and many of the attendant advantages of this invention will become more readily appreciated as the same becomes better understood by reference to the following detailed description, when taken in conjunction with the accompanying drawings, wherein:
FIG. 1 (Prior Art) is a schematic block diagram illustrating the functional block employed for MPEGT/audio coding;
FIG. 2 (Prior Art) is a functional block diagram showing the concatenation of the magnitude of a fast Fourier transform (FFT) signal to map the signal into the frequency axis;
FIG. 3 is a schematic block diagram showing a speech vector derived from nine subbands using an analysis in accord with the present invention;
FIG. 4 (prior art) is a block diagram showing a typical procedure for “training” that requires decompression of the MPEG/Audio signal;
FIG. 5 is a block diagram showing, in contrast to FIG. 4, that the present invention does not require decompression of the MPEG/Audio signal;
FIG. 6 is a graph of accuracy of recognition at various bit rates, where training occurred at two different bit rates;
FIG. 7A is an overview block diagram showing the functional blocks employed by the present invention for training;
FIG. 7B is an overview block diagram showing the functional blocks employed by the present invention for recognition;
FIGS. 8A-8D are a flow chart showing the details of the logic employed in the present invention; and
FIG. 9 is a block diagram illustrating how speech vectors comprising successive frames form a feature file.
A fundamental concept implemented in a preferred embodiment of the present invention is to take advantage of work that has gone into encoding raw PCM files into MPEG/Audio files. As illustrated in blocks 100, 102, and 104 of FIG. 7A, the training phase of speech recognition of MPEG/Audio files using speech features in feature files 103 can be achieved without decompression of the MPEG/Audio and MPEG/System files. Similarly, the recognition of speech as indicated in a block 106 of FIG. 7B, using feature files 105 can be done without decompression of MPEG/Audio or MPEG/System files 101, yielding an output of spoken words and enabling speaker identification, as indicated in a block 108. There is no need for decompression of MPEG files if speech features that are needed for speech recognition can be derived from the extracted information included in MPEG/Audio files. Once the speech features have been derived, they are first used to train Hidden Markov models that are then used to recognize speech features from unknown MPEG/Audio files 101.
The definition of MPEG/Audio can be found in the International Standard ISO/IEC Specification 11172-3, and papers describing it in detail are readily available. A block diagram 10 of the MPEG/Audio coding scheme is shown in FIG. 1. As shown therein, a signal conveying data from a PCM file 12 is input to a filter bank 14 that implements a time/frequency mapping, and also to a psychoacoustic model 16. Unlike speech-specific coders, MPEG/Audio does not assume anything about the source. Psychoacoustically unimportant information are thrown out or are coded with less bits by the psychoacoustic model. While the compression is lossy, perceptually, the decompressed signal is the same, even at six-to-one compression.
There are three different layers that can be used for the encoding process illustrated in FIG. 1; however, common to all is filter bank 14 that performs the time/frequency mapping of the signal into 32 equal-width subbands. A 512-sample moving window is used in the analysis and operates on 32 samples at a time. In a block 18, bit allocation, quantization, and coding are implemented. The output of each subband is decimated by 32 so that the number of output samples equal the number of input samples for each time unit. The output of the subbands are coded in block 18, based on the information received from either one of two psychoacoustic models 16. The output of each subband can be described with the following equations:
where st [i] is the output of the ith subband at time t, x[t] is the audio sample at time t, and Hi[n] is defined as:
where h[n] is a prototype lowpass filter. In this form, it is evident that each subband has its own bandpass filter as a result of the cosine term shifting the lowpass response h[n] to the appropriate frequency band. The center frequency for each of the subband filters is at odd multiples of
where T is the sampling period and the bandwidth of each filter is
A problem with the filters is that they are not sharp and aliasing can occur between adjacent subbands. For example, a signal at a single frequency can have effect on two subbands, which can then affect compression performance. The bitstream formatting in a block 20 involves decompression to resynthesize the subband samples back into PCM using interpolation and requires complicated filters to remove the aliasing. The output is an encoded bitstream 20.
Hidden Markov Models
There are many techniques that have been used for speech recognition. Of all the techniques, Hidden Markov models (HMMs) have to date yielded the best performance of all the techniques, since they are best at capturing and modeling the variability in speech. HMMs can represent events that have statistics that change over time, and this characteristic is particularly important for speech recognition, since it can account for variations in speech (i.e., speaking slowly and quickly).
Another advantage of HMMs is that the training data does not need to be labeled since word or phone boundaries are automatically absorbed into the model during training. The parameters of the HMMs can be determined entirely by computation, thus making them relatively easy to train. HMMs can be used for isolated as well as continuous word recognition, and they can be used to model at the subword, word, or even sentence levels. The HMM approach is very different from template matching and dynamic time warping techniques, which try to code temporal and speaker variations in a brute-force way.
Features Derived from MPEG/Audio for Speech Recognition
One way to perform speech recognition is to use a filter bank analysis. The raw PCM samples are broken up into short overlapping analysis segments that are used to compute the corresponding speech vector for training and recognition. The frequency content of each segment is determined by using a fast Fourier transform (FFT) as part of the process. In MPEG/Audio, the original audio has already been mapped into the frequency domain through the different subbands. The lower subbands contain low frequencies and the higher subbands contain higher frequencies. Unfortunately, synthesizing an exact copy of the FFT for the small overlapping segments in the original audio from the subband signals is complicated. However, a reasonable speech feature can be derived for both training and recognition by using only the subband signals.
An approach used in the present invention is to take the magnitude 34 of an FFT 32 of each subband 30 and construct an overall vector by concatenating each magnitude, generally as shown in FIG. 2. For purposes of illustration, it is assumed that the original audio was sampled at 32 kHz. Therefore, each subband 30 has a bandwidth of 500 Hz. Assuming that most human speech lies below 4.5 kHz, only subbands zero to eight need be considered. Note that the subbands that are considered can vary depending on the assumptions made. Taking magnitude 34 of FFTs 32 and concatenating them (mapping them into a frequency axis 40) amounts to shifting magnitudes 38 into appropriate frequency values 42 for the corresponding subband as shown in the graph portion of FIG. 2.
In computing the speech vectors to be used for training and recognition in an exemplary application of the present invention, a 25 ms analysis window was used, with 15 ms overlap between windows. At 32 kHz, this approach yielded 800 samples for the analysis window. Since the subband signals are decimated by 32, the number of samples dropped to 25 for each subband.
FIGS. 8A through 8D shows the logical steps of the process. Following a Start block in FIG. 8A, the logic begins with a block 110, in which an MPEG/Audio file encoded using layers I or II is input for processing. A block 112 provides that subband samples are extracted from the MPEG/Audio files. In the example, the subband samples are framed into 25-sample windows for all of the m subbands in a step 114, where the variable m is equal to nine. This processing step is repeated until there are no more subband samples.
FIG. 8B shows the steps that are applied to one of the m subbands. A decision block 116 determines if enough samples have been obtained for one frame. If not, the process is concluded and terminates. However, if the response to decision block 116 is affirmative, the logic continues with a block 118 in which the samples are framed in k samples. The frame size s is typically 20-50 msec. and the number of samples k in the frame is defined as indicated in the text and as illustrated at 120 in the Figure. In each subband analysis as noted in a block 122, the samples are windowed using a Hamming window 124 to counter the edge effects. In a block 126, the samples are then zero padded to a power of 2, as illustrated at 128, to compute the FFT, as provided in a block 130. Finally, in a block 132, the magnitude is taken, and as shown in a block 136 in FIG. 8C, the magnitudes 138 a, 138 b, . . . 138 m of the various subbands are concatenated into a single frequency vector 140. Once mapped into the frequency axis, this frequency vector gives a reasonable starting point for derivation of the speech vectors.
At this point, the processing follows established filterbank analysis procedure. In a block 142, the frequency vector is mel-warped into a mel-scale that is defined by the equation:
where f is frequency. The warping is achieved by multiplying the magnitudes in the frequency vector (i.e., in the corresponding frequency axis) with triangular filters 144 corresponding to the mel-scale. In this example, 26 triangular filters 144 were used, and they correspond to the outputs of the filter bank, as indicated at 146. Next, in a block 148, cepstral coefficients are computed from the output of the filter bank using the discrete cosine transform:
where i is the ith cepstral coefficient and N is the number of filter bank channels. Twelve cepstral coefficients plus the zeroth c0 cepstral coefficient (the energy component) were computed in the example. These thirteen coefficients make up the so-called static coefficients. Better performance can be gained by adding time derivatives to the basic static coefficients. The delta coefficients are computed from the static coefficients, as indicated in a block 150, using the following regression formula:
where dt is the delta coefficient at time t and w is set to two for the number of delta windows.
When the preceding formula is used at the beginning of this step, the static coefficients of the first speech vector is used for ct−1 and ct−2 and similarly for the end. Also, as noted in a block 152, the acceleration coefficients are computed, that is, the delta of the deltas and Equation 5 is applied to the deltas to obtain the acceleration coefficients.
As shown in FIG. 3, a total of 39 coefficients are generated for each speech vector. These coefficients are broken down into thirteen static coefficients 52, thirteen delta coefficients 54, and thirteen acceleration coefficients 56. Note that in a block 154 of FIG. 8D, there is only one speech vector 50 that is computed for all nine subbands per analysis. The resulting speech vectors 50 for each frame, which were derived from the MPEG/Audio files, comprise a feature file 162, as illustrated in FIG. 9. These speech vectors are input into the HMM for both training and recognition, as described in greater detail below.
Training Recognition, and Results
The HMM-based speech recognizer that was used in the preferred embodiment of the present invention was the HTK package obtained from Entropic Cambridge Research Laboratory. To train the HMM, a grammar of the allowed sentences is first created. The test grammar is a simple “telephone dialer” that consist of digits, proper names, and some verbs, giving a total of 17 unique words. Some examples of valid sentences are “DIAL SIX SEVEN TWO ONE” and “CALL JILL”. For training, one speaker recorded 200 continuously spoken sentences that were randomly generated from the grammar, giving a total of 1,122 words for the set.
Similarly for recognition, 300 sentences spoken continuously with different styles (e.g., quickly, slowly) were recorded by the same speaker, giving a total of 2,018 words for the set (see Table 1). Most of the sentences in the training set are not found in the recognition set and vice versa. Once recorded, the PCM files are converted into MPEG/Audio layers I and II.
Training and Recognition Sets (Recorded by One Speaker)
Continuously (Fast and Slow)
The training set was recorded at a different time than the recognition set. The sentences were sampled at 32 kHz, mono, into 16-bit PCM files using the built-in audio facility on an HP 9000/C110 computer. The amount of compression in MPEG/Audio can be tuned by encoding the files into one of several different output bit rates. For high compression, a low bit rate can be used, and similarly, for low compression, a high bit rate can be used.
For training, the PCM files are converted into MPEG/Audio layers I and II files at various output bit rates. A program using the technique described above was used to read the MPEG/Audio files and generate the corresponding speech vectors that was used to train the recognizer. For recognition, the recognition set is converted into MPEG/Audio layers I and II, also at various output bit rates. The MPEG/Audio files are read and the corresponding derived speech vectors are fed into the recognizer for recognition. The recognizer has facilities to determine the accuracy of each run, and these numbers are used for the measurement.
Training the HMM
Once the grammar has been created, a pronunciation dictionary is used to break up each word into its individual sub-word monophones. For example, the word CALL is broken into three monophones: k, ao, l, which when uttered in sequence, will produce the sound for the entire word. No segmentation or labeling of the utterances was necessary during training or recognition.
Each monophone is assigned an HMM that will be trained. The topology of the HMMs is a simple three-state, with a short pause model added to the center. In the process, the monophones are eventually converted into tied-state triphones. In each step, several invocations of parameter re-estimation are done until the final models are complete and ready for use for recognition. Complete details on training of the HMMs can be found in the HTK manual.
Training and Recognizing from MPEG/Audio
A typical prior art training procedure that requires decompression of MPEG/Audio files is illustrated in FIG. 4. Each MPEG/Audio file 60 must first be decompressed into PCM 64 in a block 62, then HTK's HCopy tool 66 is used to compute the corresponding feature file. The feature file consists of a sequence of speech vectors of mel-warped cepstral, delta, and acceleration coefficients (MCDA) files 68. Once generated, the MCDA files are used by HTK 70 to estimate the parameters for the HMM models.
However, using the technique of the present invention 72, as described above, the steps required for training reduces to those shown in FIG. 5. An intermediate PCM file 64 is not necessary. Instead, the subband signals are extracted directly from MPEG/Audio files 60 and are used to compute the MCDA files 68. Note that for the same MPEG/Audio file, the MCDA file from FIG. 4 is not the same as the MCDA file from FIG. 5. For recognition, the exact same procedure shown in FIG. 5 is used since HTK can be used to recognize MCDA files. Using HTK's HVite and HResults tools, the accuracy of the recognition can be determined. The sentences used to test for recognition accuracy are also taken from the “recognition set” (see Table 1).
For comparison, raw PCM versions of the sets in Table 1 were used for training and recognition. The best result that was obtained is shown in Table 2, where a 99.65% accuracy was achieved. Here, accuracy is defined as:
Results Using Raw PCM for Both Training and Recognition
In Table 2, the Words column refers to the total number of words, Deletion refers to the number of incorrect deletions, Substitution refers to the number of incorrect substitutions, and Insertion refers to the number of incorrect insertions. Thus, all results obtained will follow the procedure shown in FIG. 5.
Results Using Different Output Bit Rates
The accuracy resulting from training using one output bit rate of MPEG/Audio layer I files and recognition at various output bit rates is shown in Table 3. The plot in FIG. 6 shows the accuracies for training at 32 kbit/s and 192 kbit/s and recognition using various output bit rates. The result shows that the accuracies are essentially the same for training and recognition of MPEG/Audio layer I files with bit rates above 32 kbit/s. Severe distortions are present when coding at 32 kbit/s and as expected, poor recognition results were obtained using models that were trained above 32 kbit/s.
Training At One Bit Rate & Recognition At Other Bit Rates (Layer I)
Recognition at kbit/s
However, it is possible to train the HMM using files coded at 32 kbit/s and achieve good recognition for files coded at the same bit rate. Unfortunately, performance drops when recognizing files coded at higher bit rates as shown in a graph 80, in FIG. 6. In this graph, the squares indicate the accuracy where the training was at the higher output bit rate of 192 kbit/s, while the circles indicate the recognition accuracy for the case where the training was at the lower bit rate of 32 kbit/s. Here, the distortions in lower bit rate files were absorbed by the HMM models, but these distortions are not found in the higher bit rate files.
Recognition Accuracy For 32 Kbit/S Training (Layer II)
Recognition at kbit/s
Table 4 shows an improvement in performance by keeping everything the same except for coding the training and recognition sets using layer II. Here, training at 32 kbit/s did not degrade the recognition of files coded at higher bit rates. This is to be expected since layer II can compress more information at the same bit rate than layer I can.
Results from Interlayer Recognition
Recognition of one layer is not limited to the layer on which it was trained. The recognizer can be trained on layer I files and be used to recognize layer II files and vice versa. At an output bit rate of 96 kbit/s, the interlayer recognition accuracy is quite good. Given training and recognition at an appropriate bit rate, Table 5 shows that interlayer recognition is possible.
Interlayer Recognition for Training & Recognition at 96 kbit/s
Recognition Results from Varying the Number of Subbands
For this subband experiment, the subbands that were used in creating the speech vectors are varied. For training, nine subbands are used, that is, subbands 0 to 8. The result of varying the subbands (i.e., so that there is a mismatch between the subbands used for training and for recognition) on recognition is shown in Table 6. Note the drop in accuracy if the subbands considered are from 0 to 9, i.e., when the mismatch occurs because more subbands were considered during recognition than in training.
Accuracy Based on the Subbands Considered
0 → 0
0 → 1
0 → 2
0 → 3
0 → 4
0 → 5
0 → 6
0 → 7
0 → 8
0 → 9
The experiments have shown that the technique employed in the present invention for deriving speech features from only the subband signals of MPEG/Audio, combined with HMMs, can result in good performance. Training and recognition is possible using different output bit rates. In cases where there are distortion, the HMMs can be trained to absorb the distortions and still perform remarkably well. Interlayer recognition is possible so that training is not confined to one layer, thereby making the technique flexible. With better compression, layer II consistently produced better results than layer I in both training and recognition.
In this method, the choice of the subbands to be considered is important. As shown in Table 6, the subbands should be exactly the same for both training and recognition. For example, it is possible to use only subband 0 for both training and recognition. However, the speech vectors may not be sufficiently decorralated to produce reasonable accuracy. Also, for good recognition accuracy, the subband used should correspond to the range in which human speech is found. Therefore, once the appropriate subband has been determined, it is important to use it consistently for both training and recognition. to avoid the mismatch problem noted above.
A novel technique for speech recognition in the compressed domain of MPEG/Audio has been described above, and experiments have confirmed that the recognition performance of the present invention is as good as training and recognition in the uncompressed domain. The present invention provides a technique that is fast and is applicable to both layers I and II of MPEG/Audio. When this technique is combined with an HMM-based speech recognizer, distortions that are present in MPEG/Audio can be absorbed during training and still yield good performance during recognition of similarly distorted files. Another advantage of using HMMs is that a crude form of speaker independence can be achieved simply by using samples from different speakers during training.
Although the present invention has been described in connection with the preferred form of practicing it, those of ordinary skill in the art will understand that many modifications can be made thereto within the scope of the claims that follow. Accordingly, it is not intended that the scope of the invention in any way be limited by the above description, but instead be determined entirely by reference to the claims that follow.
|Patente citada||Fecha de presentación||Fecha de publicación||Solicitante||Título|
|US4866777 *||9 Nov 1984||12 Sep 1989||Alcatel Usa Corporation||Apparatus for extracting features from a speech signal|
|US5305422 *||28 Feb 1992||19 Abr 1994||Panasonic Technologies, Inc.||Method for determining boundaries of isolated words within a speech signal|
|US6003004 *||8 Ene 1998||14 Dic 1999||Advanced Recognition Technologies, Inc.||Speech recognition method and system using compressed speech data|
|1||*||Proceedings., IEEE International Conference on Multimedia Computing and Systems '97. Yapp et al., "Speech REcognition on MPEG/Audio encoded files", pp624-625. Jun. 3-6, 1997.|
|Patente citante||Fecha de presentación||Fecha de publicación||Solicitante||Título|
|US6542869 *||11 May 2000||1 Abr 2003||Fuji Xerox Co., Ltd.||Method for automatic analysis of audio including music and speech|
|US6603921 *||1 Jul 1998||5 Ago 2003||International Business Machines Corporation||Audio/video archive system and method for automatic indexing and searching|
|US6681207 *||12 Ene 2001||20 Ene 2004||Qualcomm Incorporated||System and method for lossy compression of voice recognition models|
|US6701291 *||2 Abr 2001||2 Mar 2004||Lucent Technologies Inc.||Automatic speech recognition with psychoacoustically-based feature extraction, using easily-tunable single-shape filters along logarithmic-frequency axis|
|US6721698 *||27 Oct 2000||13 Abr 2004||Nokia Mobile Phones, Ltd.||Speech recognition from overlapping frequency bands with output data reduction|
|US6915246 *||17 Dic 2001||5 Jul 2005||International Business Machines Corporation||Employing speech recognition and capturing customer speech to improve customer service|
|US7028325 *||13 Sep 2000||11 Abr 2006||Microsoft Corporation||Annotating programs for automatic summary generation|
|US7058565||17 Dic 2001||6 Jun 2006||International Business Machines Corporation||Employing speech recognition and key words to improve customer service|
|US7167544 *||22 Nov 2000||23 Ene 2007||Siemens Aktiengesellschaft||Telecommunication system with error messages corresponding to speech recognition errors|
|US7203635 *||27 Jun 2002||10 Abr 2007||Microsoft Corporation||Layered models for context awareness|
|US7376556 *||2 Mar 2004||20 May 2008||Phoenix Solutions, Inc.||Method for processing speech signal features for streaming transport|
|US7403894||15 Mar 2005||22 Jul 2008||Microsoft Corporation||Annotating programs for automatic summary generations|
|US7418394 *||28 Abr 2005||26 Ago 2008||Dolby Laboratories Licensing Corporation||Method and system for operating audio encoders utilizing data from overlapping audio segments|
|US7454342 *||19 Mar 2003||18 Nov 2008||Intel Corporation||Coupled hidden Markov model (CHMM) for continuous audiovisual speech recognition|
|US7552336 *||8 Nov 2004||23 Jun 2009||Microsoft Corporation||Watermarking with covert channel and permutations|
|US7620552||4 Mar 2005||17 Nov 2009||Microsoft Corporation||Annotating programs for automatic summary generation|
|US7647225||20 Nov 2006||12 Ene 2010||Phoenix Solutions, Inc.||Adjustable resource based speech recognition system|
|US7647340||15 Jun 2001||12 Ene 2010||Sharp Laboratories Of America, Inc.||Metadata in JPEG 2000 file format|
|US7653131||2 Dic 2005||26 Ene 2010||Sharp Laboratories Of America, Inc.||Identification of replay segments|
|US7657424||3 Dic 2004||2 Feb 2010||Phoenix Solutions, Inc.||System and method for processing sentence based queries|
|US7657907||30 Sep 2002||2 Feb 2010||Sharp Laboratories Of America, Inc.||Automatic user profiling|
|US7672841||19 May 2008||2 Mar 2010||Phoenix Solutions, Inc.||Method for processing speech data for a distributed recognition system|
|US7698131||9 Abr 2007||13 Abr 2010||Phoenix Solutions, Inc.||Speech recognition system for client devices having differing computing capabilities|
|US7702508||3 Dic 2004||20 Abr 2010||Phoenix Solutions, Inc.||System and method for natural language processing of query answers|
|US7725307||29 Ago 2003||25 May 2010||Phoenix Solutions, Inc.||Query engine for processing voice based queries including semantic decoding|
|US7725320||9 Abr 2007||25 May 2010||Phoenix Solutions, Inc.||Internet based speech recognition system with dynamic grammars|
|US7725321||23 Jun 2008||25 May 2010||Phoenix Solutions, Inc.||Speech based query system using semantic decoding|
|US7729904||3 Dic 2004||1 Jun 2010||Phoenix Solutions, Inc.||Partial speech processing device and method for use in distributed systems|
|US7793205||8 Jul 2005||7 Sep 2010||Sharp Laboratories Of America, Inc.||Synchronization of video and data|
|US7831426||23 Jun 2006||9 Nov 2010||Phoenix Solutions, Inc.||Network based interactive speech recognition system|
|US7853865||8 Jul 2005||14 Dic 2010||Sharp Laboratories Of America, Inc.||Synchronization of video and data|
|US7873519||31 Oct 2007||18 Ene 2011||Phoenix Solutions, Inc.||Natural language speech lattice containing semantic variants|
|US7877774 *||19 Abr 2000||25 Ene 2011||At&T Intellectual Property Ii, L.P.||Browsing and retrieval of full broadcast-quality video|
|US7904814||13 Dic 2001||8 Mar 2011||Sharp Laboratories Of America, Inc.||System for presenting audio-video content|
|US7912702||31 Oct 2007||22 Mar 2011||Phoenix Solutions, Inc.||Statistical language model trained with semantic variants|
|US8020183||30 Mar 2001||13 Sep 2011||Sharp Laboratories Of America, Inc.||Audiovisual management system|
|US8028314 *||26 May 2000||27 Sep 2011||Sharp Laboratories Of America, Inc.||Audiovisual information management system|
|US8073684 *||25 Abr 2003||6 Dic 2011||Texas Instruments Incorporated||Apparatus and method for automatic classification/identification of similar compressed audio files|
|US8108209 *||26 May 2009||31 Ene 2012||Coding Technologies Sweden Ab||Method for reduction of aliasing introduced by spectral envelope adjustment in real-valued filterbanks|
|US8145475||27 May 2009||27 Mar 2012||Coding Technologies Sweden Ab||Method for reduction of aliasing introduced by spectral envelope adjustment in real-valued filterbanks|
|US8214741||22 May 2002||3 Jul 2012||Sharp Laboratories Of America, Inc.||Synchronization of video and data|
|US8229734||23 Jun 2008||24 Jul 2012||Phoenix Solutions, Inc.||Semantic decoding of user queries|
|US8346566||31 Ago 2010||1 Ene 2013||Dolby International Ab||Method for reduction of aliasing introduced by spectral envelope adjustment in real-valued filterbanks|
|US8352277||9 Abr 2007||8 Ene 2013||Phoenix Solutions, Inc.||Method of interacting through speech with a web-connected server|
|US8356317||13 Jun 2005||15 Ene 2013||Sharp Laboratories Of America, Inc.||Presence based technology|
|US8498876||18 Jul 2012||30 Jul 2013||Dolby International Ab|
|US8527268||30 Jun 2010||3 Sep 2013||Rovi Technologies Corporation||Method and apparatus for improving speech recognition and identifying video program material or content|
|US8606587||18 Jul 2012||10 Dic 2013||Dolby International Ab|
|US8689253||3 Mar 2006||1 Abr 2014||Sharp Laboratories Of America, Inc.||Method and system for configuring media-playing sets|
|US8761545||19 Nov 2010||24 Jun 2014||Rovi Technologies Corporation||Method and apparatus for identifying video program material or content via differential signals|
|US8762152||1 Oct 2007||24 Jun 2014||Nuance Communications, Inc.||Speech recognition system interactive agent|
|US8776142||2 Sep 2009||8 Jul 2014||Sharp Laboratories Of America, Inc.||Networked video devices|
|US8878041||27 May 2009||4 Nov 2014||Microsoft Corporation||Detecting beat information using a diverse set of correlations|
|US8949899||13 Jun 2005||3 Feb 2015||Sharp Laboratories Of America, Inc.||Collaborative recommendation system|
|US9020326||31 Oct 2005||28 Abr 2015||At&T Intellectual Property Ii, L.P.||System and method for content-based navigation of live and recorded TV and video programs|
|US9020825 *||25 Sep 2012||28 Abr 2015||Rawles Llc||Voice gestures|
|US9042703||31 Oct 2005||26 May 2015||At&T Intellectual Property Ii, L.P.||System and method for content-based navigation of live and recorded TV and video programs|
|US9076448||10 Oct 2003||7 Jul 2015||Nuance Communications, Inc.||Distributed real time speech recognition system|
|US9171545||30 Nov 2010||27 Oct 2015||At&T Intellectual Property Ii, L.P.||Browsing and retrieval of full broadcast-quality video|
|US9190063||31 Oct 2007||17 Nov 2015||Nuance Communications, Inc.||Multi-language speech recognition system|
|US9218818||27 Abr 2012||22 Dic 2015||Dolby International Ab||Efficient and scalable parametric stereo coding for low bitrate audio coding applications|
|US9401144||24 Abr 2015||26 Jul 2016||Amazon Technologies, Inc.||Voice gestures|
|US9542950||14 Nov 2013||10 Ene 2017||Dolby International Ab|
|US9741395||27 Abr 2015||22 Ago 2017||At&T Intellectual Property Ii, L.P.||System and method for content-based navigation of live and recorded TV and video programs|
|US9743144||22 May 2015||22 Ago 2017||At&T Intellectual Property Ii, L.P.||System and method for content-based navigation of live and recorded TV and video programs|
|US20010010523 *||12 Mar 2001||2 Ago 2001||Sezan M. Ibrahim||Audiovisual information management system|
|US20020078438 *||19 Mar 2001||20 Jun 2002||U.S. Philps Corporation||Video signal analysis and storage|
|US20020095297 *||17 Ene 2002||18 Jul 2002||Satoshi Hasegawa||Device and method for processing audio information|
|US20020180774 *||13 Dic 2001||5 Dic 2002||James Errico||System for presenting audio-video content|
|US20030061610 *||27 Mar 2001||27 Mar 2003||Errico James H.||Audiovisual management system|
|US20030076448 *||27 Sep 2002||24 Abr 2003||Hao Pan||Identification of replay segments|
|US20030115056 *||17 Dic 2001||19 Jun 2003||International Business Machines Corporation||Employing speech recognition and key words to improve customer service|
|US20030115064 *||17 Dic 2001||19 Jun 2003||International Business Machines Corporaton||Employing speech recognition and capturing customer speech to improve customer service|
|US20030121040 *||22 May 2002||26 Jun 2003||Ferman A. Mufit||Audiovisual management system|
|US20040002838 *||27 Jun 2002||1 Ene 2004||Oliver Nuria M.||Layered models for context awareness|
|US20040030556 *||25 Jun 2003||12 Feb 2004||Bennett Ian M.||Speech based learning/training system using semantic decoding|
|US20040073428 *||10 Oct 2002||15 Abr 2004||Igor Zlokarnik||Apparatus, methods, and programming for speech synthesis via bit manipulations of compressed database|
|US20040073918 *||30 Sep 2002||15 Abr 2004||Ferman A. Mufit||Automatic user profiling|
|US20040186718 *||19 Mar 2003||23 Sep 2004||Nefian Ara Victor||Coupled hidden markov model (CHMM) for continuous audiovisual speech recognition|
|US20040197088 *||31 Mar 2003||7 Oct 2004||Ferman Ahmet Mufit||System for presenting audio-video content|
|US20040215447 *||25 Abr 2003||28 Oct 2004||Prabindh Sundareson||Apparatus and method for automatic classification/identification of similar compressed audio files|
|US20040249635 *||2 Mar 2004||9 Dic 2004||Bennett Ian M.||Method for processing speech signal features for streaming transport|
|US20040255150 *||19 Jul 2004||16 Dic 2004||Sezan Muhammed Ibrahim||Audiovisual information management system|
|US20040261095 *||19 Jul 2004||23 Dic 2004||Sezan Muhammed Ibrahim||Audiovisual information management system|
|US20040267805 *||19 Jul 2004||30 Dic 2004||Sezan Muhammed Ibrahim||Audiovisual information management system|
|US20040268383 *||19 Jul 2004||30 Dic 2004||Sezan Muhammed Ibrahim||Audiovisual information management system|
|US20040268389 *||19 Jul 2004||30 Dic 2004||Sezan Muhammed Ibrahim||Audiovisual information management system|
|US20040268390 *||19 Jul 2004||30 Dic 2004||Muhammed Ibrahim Sezan||Audiovisual information management system|
|US20050060641 *||28 Oct 2004||17 Mar 2005||Sezan Muhammed Ibrahim||Audiovisual information management system with selective updating|
|US20050091165 *||28 Oct 2004||28 Abr 2005||Sezan Muhammed I.||Audiovisual information management system with usage preferences|
|US20050091685 *||28 Oct 2004||28 Abr 2005||Sezan Muhammed I.||Audiovisual information management system|
|US20050091686 *||28 Oct 2004||28 Abr 2005||Sezan Muhammed I.||Audiovisual information management system with seasons|
|US20050108542 *||8 Nov 2004||19 May 2005||Microsoft Corporation||Watermarking with covert channel and permutations|
|US20050120033 *||28 Oct 2004||2 Jun 2005||Sezan Muhammed I.||Audiovisual information management system with identification prescriptions|
|US20050120034 *||28 Oct 2004||2 Jun 2005||Sezan Muhammed I.||Audiovisual information management system with advertising|
|US20050131727 *||28 Oct 2004||16 Jun 2005||Sezan Muhammed I.||Audiovisual information management system with user identification|
|US20050141864 *||28 Oct 2004||30 Jun 2005||Sezan Muhammed I.||Audiovisual information management system with preferences descriptions|
|US20050159956 *||4 Mar 2005||21 Jul 2005||Microsoft Corporation||Annotating programs for automatic summary generation|
|US20050160457 *||15 Mar 2005||21 Jul 2005||Microsoft Corporation||Annotating programs for automatic summary generations|
|US20050188328 *||28 Oct 2004||25 Ago 2005||Sezan Muhammed I.||Audiovisual information management system with presentation service|
|US20050271146 *||8 Jul 2005||8 Dic 2005||Sharp Laboratories Of America, Inc.||Synchronization of video and data|
|US20050271269 *||8 Jul 2005||8 Dic 2005||Sharp Laboratories Of America, Inc.||Synchronization of video and data|
|US20060083304 *||2 Dic 2005||20 Abr 2006||Sharp Laboratories Of America, Inc.||Identification of replay segments|
|US20060174277 *||13 Jun 2005||3 Ago 2006||Sezan M I||Networked video devices|
|US20060247928 *||28 Abr 2005||2 Nov 2006||James Stuart Jeremy Cowdery||Method and system for operating audio encoders in parallel|
|US20060282856 *||13 Jun 2005||14 Dic 2006||Sharp Laboratories Of America, Inc.||Collaborative recommendation system|
|US20070050827 *||31 Oct 2005||1 Mar 2007||At&T Corp.||System and method for content-based navigation of live and recorded TV and video programs|
|US20070098350 *||31 Oct 2005||3 May 2007||At&T Corp.||System and method for content-based navigation of live and recorded TV and Video programs|
|US20070185717 *||9 Abr 2007||9 Ago 2007||Bennett Ian M||Method of interacting through speech with a web-connected server|
|US20080215327 *||19 May 2008||4 Sep 2008||Bennett Ian M||Method For Processing Speech Data For A Distributed Recognition System|
|US20090234646 *||27 May 2009||17 Sep 2009||Kristofer Kjorling||Method for Reduction of Aliasing Introduced by Spectral Envelope Adjustment in Real-Valued Filterbanks|
|US20090259479 *||26 May 2009||15 Oct 2009||Coding Technologies Sweden Ab|
|US20090317052 *||2 Sep 2009||24 Dic 2009||Sharp Laboratories Of America, Inc.||Networked video devices|
|US20100300271 *||27 May 2009||2 Dic 2010||Microsoft Corporation||Detecting Beat Information Using a Diverse Set of Correlations|
|US20110054914 *||31 Ago 2010||3 Mar 2011||Kristofer Kjoerling||Method for Reduction of Aliasing Introduced by Spectral Envelope Adjustment in Real-Valued Filterbanks|
|US20110072466 *||30 Nov 2010||24 Mar 2011||At&T Intellectual Property Ii, L.P. Via Transfer From At&T Corp.||Browsing and Retrieval of Full Broadcast-Quality Video|
|US20150302869 *||16 Ene 2015||22 Oct 2015||Arthur Charles Tomlin||Conversation, presence and context detection for hologram suppression|
|WO2012003125A1 *||22 Jun 2011||5 Ene 2012||Rovi Technologies Corporation||Method and apparatus for identifying video program material or content via frequency translation or modulation schemes|
|WO2012012074A1 *||22 Jun 2011||26 Ene 2012||Rovi Technologies Corporation||Method and apparatus for identifying video program material or content via nonlinear transformations|
|Clasificación de EE.UU.||704/251, 704/256, 704/E15.004, 704/E15.028, 704/200.1|
|Clasificación internacional||G10L15/02, G10L15/14|
|Clasificación cooperativa||G10L15/02, G10L15/142|
|Clasificación europea||G10L15/02, G10L15/14M|
|25 Oct 1998||AS||Assignment|
Owner name: WASHINGTON, UNIVERSITY OF, WASHINGTON
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:YAPP, LAWRENCE;REEL/FRAME:009597/0920
Effective date: 19981021
|26 Oct 1998||AS||Assignment|
Owner name: WASHINGTON, UNIVERSITY OF, WASHINGTON
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:ZICK, GREGORY L.;REEL/FRAME:009597/0903
Effective date: 19980731
|12 Oct 2005||SULP||Surcharge for late payment|
|12 Oct 2005||FPAY||Fee payment|
Year of fee payment: 4
|10 Nov 2009||FPAY||Fee payment|
Year of fee payment: 8
|10 Nov 2009||SULP||Surcharge for late payment|
Year of fee payment: 7
|15 Nov 2013||REMI||Maintenance fee reminder mailed|
|9 Abr 2014||LAPS||Lapse for failure to pay maintenance fees|
|27 May 2014||FP||Expired due to failure to pay maintenance fee|
Effective date: 20140409